On Thu, 2011-04-28 at 16:08 +0100, Mel Gorman wrote: > On Thu, Apr 28, 2011 at 09:25:14AM -0500, James Bottomley wrote: > > On Thu, 2011-04-28 at 15:07 +0100, Mel Gorman wrote: > > > On Thu, Apr 28, 2011 at 03:52:28PM +0200, Jan Kara wrote: > > > > On Thu 28-04-11 12:36:30, Colin Ian King wrote: > > > > > One more data point to add, I've been looking at an identical issue when > > > > > copying large amounts of data. I bisected this - and the lockups occur > > > > > with commit > > > > > 3e7d344970673c5334cf7b5bb27c8c0942b06126 - before that I don't see the > > > > > issue. With this commit, my file copy test locks up after ~8-10 > > > > > iterations, before this commit I can copy > 100 times and don't see the > > > > > lockup. > > > > Adding Mel to CC, I guess he'll be interested. Mel, it seems this commit > > > > of yours causes kswapd on non-preempt kernels spin for a *long* time... > > > > > > > > > > I'm still thinking about the traces which do not point the finger > > > directly at compaction per-se but it's possible that the change means > > > kswapd is not reclaiming like it should be. > > > > > > To test this theory, does applying > > > [d527caf2: mm: compaction: prevent kswapd compacting memory to reduce > > > CPU usage] help? > > > > I can answer definitively no to this. The upstream kernel I reproduced > > this on has that patch included. > > > > So it is. > > Another consequence of this patch is that when high order allocations > are in progress (is the test case fork heavy in any way for > example? It's a simple huge untar, so it shouldn't fork. > alternatively, it might be something in the storage stack > that requires high-order allocs) I've tried switching from CFQ to deadline with no apparent changes > we are no longer necessarily going > to sleep because of should_reclaim_continue() check. This could > explain kswapd-at-99% but would only apply if CONFIG_COMPACTION is > set (does unsetting CONFIG_COMPACTION help). If the bug only triggers > for CONFIG_COMPACTION, does the following *untested* patch help any? Turning off COMPACTION and HUGEPAGES doesn't help ... kswapd still goes to 99% on the PREEMPT kernel, so it doesn't seem to be related James -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
